preface
Collected some open automotive ReID data sets for your reference.
VeRi776
Containing more than 50,000 images of 776 vehicles taken by 20 cameras covering an area of 1.0 square kilometers over 24 hours, this data set is scalable enough to be used for vehicle ReID and other related studies. Images are captured in real-world unconstrained monitoring scenarios and labeled with different properties such as: BBox, type, color, and brand. Thus the complex model of the vehicle ReID can be learned and evaluated. Each vehicle is photographed by 2 to 18 cameras at different viewpoints, lighting, resolution and occlusion, providing a high recurrence rate for the vehicle ReID in the actual monitoring environment. It is also marked with enough license plate and temporal information, such as the BBox of the plate, the slats, the vehicle’s time stamp and the distance between the adjacent cameras.
Download address: vehiclereid. Making. IO/VeRi /
VehicleID
Contains data captured during the day from multiple real surveillance cameras spread across a small Chinese city. There are 26,267 vehicles in the entire data set (221,763 images in total). Each image carries an ID tag that corresponds to a real-world identity. In addition, vehicle model information was manually marked for 10,319 vehicles (90,196 images in total).
Download address: www.pkuml.org/resources/p…
VERI-Wild
The vehicle images were captured by a CCTV system with 174 cameras covering more than 200 square kilometers of the city. The camera is shooting 24 hours a day for 30 days, and its long run takes into account the vehicle’s real weather and light problems. It contains 400,000 images, 40,000 vehicle tags. The data set provides camera IDS, time stamps, and tracking relationships between cameras.
Project address: github.com/PKU-IMRE/VE…
N-CARS
The data set is based on a real-world event data set consisting of approximately 24,000 samples taken from vehicle driving in urban and highway environments. The ATIS camera mounted behind the windscreen of the car was used to capture 80 minutes and convert it into a conventional grayscale image and mark the sample. The data set consists of 12336 automobile samples and 11693 non-automobile samples. The training set was divided into 7940 automobile samples and 7482 background samples, and the test set included 4396 automobile samples and 4211 background test samples.
Download at www.prophesee.ai/dataset-n-c…
PKU-VD
The dataset contains two large vehicle datasets (VD1 and VD2) that take images from real-world unrestricted scenes in two cities, respectively. VD1 is obtained from the high-resolution traffic camera, and the image in VD2 is obtained from the surveillance video. The authors perform vehicle detection on the original data to ensure that each image contains only one vehicle. Due to privacy restrictions, all license plate numbers have been covered in black. All vehicle images are taken from the front view. The dataset provides a variety of attribute annotations for each image, including identity numbers, accurate vehicle models and vehicle colors. Specifically, the ID number (ID) is unique and all images belonging to the same vehicle have the same ID(ensure that there are at least two images in the data set for each vehicle ID). Provides the most accurate model types, including detailed vehicle types and different production years. For example, Audi A6L-2012&2015, Audi A6-2004, Audi A4-2006&2008 and Audi A4-2004&2005 are four different vehicle models in the data set. As for color information, 11 common colors are labeled in the dataset. The details are as follows: VD1: Originally contains 1097649 images, 1232 vehicle models, 11 vehicle colors, but delete the images with multiple vehicles and the images taken from behind the vehicles, the data set is left with 846358 images, 141756 vehicles. VD2: Originally contained 807,260 images, 79,763 vehicles, 1112 vehicle models, 11 vehicle colors. 690518 images after the same operation with VD1.
Project address: pkuml.org/resources/p…
Baidu Apollo 3D vehicle data set
In the data set, the camera folder stores the camera parameters, the CAR_Model folder stores the car model set in PKL, the CA_Poses folder stores the position of the car in the picture, the Imags folder stores the car image, and the Split folder stores the index of the training and verification images.
Download from apolloscape.auto/car_instanc…
Cars Dataset
The Cars Dataset contains 16,185 images of 196 types of automobiles. The data were divided into 8144 training images and 8041 test images, among which each category was roughly divided into 50-50. Usually categories are labeled as Make, Model, and Year levels, such as 2012 Tesla Model S or 2012 BMW M3 coupe.
Project link: ai.stanford.edu/~jkrause/ca…
CompCars
The CompCars dataset contains images from both the network and surveillance approaches. The network image is obtained by automobile collection forum, public website and other ways; Surveillance images are collected by surveillance cameras. Vehicle types only include cars and SUVs. Network images include 163 vehicle brands and 1,716 models, with a total of 136,726 images of the entire car and 27,618 images of auto parts. The image of the complete car is marked with bounding boxes and viewpoints. The monitoring nature data consists of 50,000 images of cars taken in front view angles, each of which is annotated with this model and color. Each model is labeled with five attributes, including maximum speed, displacement, number of doors, number of seats and type of car. For each vehicle model, five shooting angles are also marked, including front, back, side, front and back.
Connection of the project: mmlab.ie.cuhk.edu.hk/datasets/co…